Why is it not possible to pass matrix parameters as argument to GPU arrayfun and what to do?

Background

1. The problem can be separated into a large number of independent sub-problems.
2. All sub-problems share the same matrix parameters of significant sizes.
3. Each sub-problem performs an operation involving the matrix parameters that is hard to express with linear algebra.
4. The goal is to process the sub-problems in paralell on the GPU.

Answer 1

[Edric Ellis]

It would be great to see a simplified example of what you're trying to do. What you can do with gpuArray/arrayfun is write a nested function to access constant data from a parent workspace, a bit like this:

function [r,c] = example
% up-level variables that we will access inside the nested
% function "iFindLocation"
mat = gpuArray(magic(4));
[m,n] = size(mat);

[r,c] = arrayfun(@iFindLocation, gpuArray(1:100));
    function [r,c] = iFindLocation(in)
        r = NaN; c = NaN;
        for i = 1:m
            for j = 1:n
                % Access the matrix "mat" from the parent
                % workspace
                if mat(i,j) == in
                    r = i;
                    c = j;
                end
            end
        end
    end
end

[Anders]

99% of the code used for this project can be regarded as a black box function analogous to the function iFindLocation.

Your example code represents this problem nicely.

Follow-up question

For arrayfun to work in paralell, does it need to store copies of the matrix mat in GPU memory or is it enougth with just the one?

[Edric Ellis]

I believe the constant data used inside the nested function here does not need to be duplicated.